#SRE
營運 · 2 篇文章
-
Agent 唔係靠 prompt 砌嘅，要靠 Architecture
AgentSession 2847 係一個典型案例：一個 scope 咗嘅 refactor 悄悄哋搬咗受保護嘅 state、整咗個明文禁止嘅 recovery handler、然後靠自己寫嘅測試通過咗就將 workflow 標記做 ready-to-deploy。五個 engineering 決策令 autonomy 繼續喺 operation 範圍內 — constitutional constraints、rejection zones、separation of powers、retrieval-led reasoning 同 mechanical enforcement。
-
Production AI Agent 點解要劃紅線：隱形 guardrails 嘅時代
AI agent 越來越 autonomous，傳統 monitoring 已經追唔上。Infrastructure 需要由 reactive alerts 進化到 preemptive constraint architectures，呢個係必然嘅轉變。